An AI teaching assistant that tutors students at midnight from the instructor's own pages, cited, and never gives a graded answer
Help at midnight from the instructor's own pages, and never a graded answer.
See one case, screen by screen ↓An AI teaching assistant helps students with their coursework when the instructor is not available, ideally from the instructor's own materials and within the instructor's rules. AI Teaching Assistant gives students cited answers from their own course materials: a Tutor agent replies as two hints and a worked example, each cited to a page, a separate Guardrail agent checks every rung against sealed answer keys, and each morning the instructor sees what her sections are stuck on and approves what to re-teach.
AI in education: a buyer's guide to agentic AI for colleges, universities and school districts →
Updated
An AI teaching assistant matters because students study when help is closed. Office hours, the tutoring center and the learning assistants work days; the questions come at night. A student stuck on a lab question searches the web, pastes the graded question into a public chatbot, or waits two days. The chatbot answers from the open web, in English, cites nothing a student can check, and hands over the answer instead of the learning. A student who writes in Spanish or Vietnamese gets English-only help, or none.
The cost lands on everyone else. A learning assistant answers the same question at the lab bench eight times without knowing it is the same question, and the instructor learns that a quarter of the lecture holds one misconception from the lab reports or the midterm, weeks after a 10-minute fix would have worked. A student who writes something alarming at 1 a.m. reaches no one. And because tools are bought instructor by instructor, the college has no floor of rules, no two-person change and no record a reviewer can read. AI Teaching Assistant puts seven agents in front of that work, and the instructor decides.
Estimated: a typical gateway course, by hand and with the agents, our arithmetic from the way the work runs today, kept modest. Knowing whether a re-teach worked still takes a week of questions with the agents; without them it waits for the next exam.
A Tutor answers from the course's pages, a Guardrail agent checks every rung against sealed keys; the instructor decides what to re-teach.
Brandvale Community College, Fall 2026, week 9. BIO 141 General Biology I has one 142-student lecture and four lab sections, and Lab 4, Diffusion and osmosis, has graded post-lab questions due Monday. Here is one student's question and what the course AI assistant did with it, screen by screen, in the working solution.
Lucía types “¿Por qué la sal hace que la célula de la cebolla se encoja?” inside the learning management system. The Language agent detects Spanish and keeps the English beside it for her instructor, at translation confidence 0.97. The Guardrail agent finds the question matches sealed Lab 4 post-lab question 2 at 0.91, so the reply will be hints only, and any worked example must use a different case.
Graded Lab 4 post-lab question 2, matched 0.91: the answer is never shown.
Hint 1 asks what crosses the membrane in osmosis, the salt or the water, and cites Lab 4 handout p. 2 ¶2; the Guardrail agent checked it first: no graded answer, reading level 8.4 against the course limit of 10, on topic. Lucía tries “¿La sal entra en la célula?”, the misconception, and Hint 2 points her to Figure 4.2 on p. 3. She writes “Hacia afuera… ¿entonces el agua sale de la célula?”, and because she has tried, the worked example opens.
The Tutor agent's first draft of the worked example used the onion cell, the graded case, so the Guardrail agent sent it back. The rewrite works a red blood cell in 3% saline in three steps, from Lecture 7 slide 14. At 11:53 p.m. Lucía writes “¡Gracias! Ya lo puedo escribir.” No graded answer was shown.
Now apply the same three steps to your onion cell, in your own words.
64 questions came from 41 of the course's 142 students overnight, 21 in Spanish and 3 in Vietnamese, and the Guardrail agent rewrote 4 drafts before anyone saw them. Across the week, 13 replies were rewritten or blocked. Where the graded item has closed, such as a Lab 2 quiz question closed October 5, the instructor can override a rewrite, with a reason code.
The Insight agent groups the last 7 days, 486 questions from 118 of 142 students, by what students get wrong. At the top: 37 of 142 students, in 58 questions, mixed up the direction of osmosis, with L01 at 11, L02 at 9, L03 at 10 and L04 at 7. Each item links to the questions behind it and carries the agent's proposed action.
Adaeze reads Lucía's question beside its English, then flips the same screen and opens the questions behind the top item: Lucía M. in Spanish, Diego A. asking whether water moves toward less salt, Minh T. in Vietnamese and Fatima N. in Arabic, each with the English beside it. She sees the misconception exactly as her students wrote it.
The plan for the 8:00 a.m. lecture starts from Figure 4.2 in the Lab 4 handout, works the red blood cell example from Lecture 7 slides 12–15, and ends with two ungraded check questions and a one-line exit question. The check questions were checked against the sealed keys too; the highest match is 0.22. Every step is cited, and the evidence sits beside it.
Adaeze approves the plan for the 8:00 a.m. lecture. The agents add it to her lecture notes in the learning management system, assign the lab follow-up to Priya Natarajan (L01 and L02, 20 students) for Thursday and Tomás Ferreira (L03 and L04, 17 students) for Friday, and set the measure: osmosis-direction questions from October 21 to 27 against the baseline of 58.
Last week's re-teach on diffusion vs. osmosis cut repeat questions from 44 to 12.
Priya sees only L01 and L02, with first names and last initials and no grades or counseling alerts. The direction of osmosis is due Thursday at 9:00 a.m., with a talking point the agent drafted from the course materials; a note she drafts for students waits for Adaeze to approve. Two Spanish replies at translation confidence 0.87 and 0.89 wait for her bilingual check.
A CHEM 121 student wrote words of hopelessness. The Safety agent scored it 0.88, showed the student the 988 Suicide & Crisis Lifeline and the after-hours counseling line at once, and alerted the on-call counselor with the flagged exchange only. Marcus acknowledged at 1:18 a.m., six minutes later, and reached the student by phone at 1:24 a.m. The instructor is told only that a referral happened.
The same night's questions, seen by the five people who carry them: what their week looked like, and what it looks like now.
A Tutor answers from the course's pages, a Guardrail agent checks every rung against sealed keys; the instructor decides what to re-teach.
Answers only from the course's uploaded materials (syllabus, lab handouts, slides and readings) as Hint 1, Hint 2 and a worked example on a different case, each cited to a page, in the student's language.
Checks every draft rung against the sealed answer keys, the reading level, topic limits and exam windows before it is shown; rewrites or blocks it; replays a proposed rule change against last week's questions.
Detects the student's language and replies in it, keeps the English beside it for staff, and back-translates to check.
Scores every message for self-harm, harassment and threat signals, shows the student crisis resources at once, and alerts the on-call counselor under the behavioral intervention team protocol.
Groups the week's questions into misconceptions per lab section, drafts a re-teach and lab talking points from the instructor's own materials, and measures whether the repeat questions fell.
Reads every upload, maps it to the syllabus weeks and topics, and finds questions with no source and the reason, such as a scanned page with no text.
Builds the college review pack: questions, rungs, checks, alert times, approvals and versions, with names replaced by student IDs, and seals it.
Approve, acknowledge, co-sign rules. The agents propose; a named person decides.
An instructor can ask about any question, student, section, rule or alert. The assistant answers from the course's materials and records, with the source beside every number, and opens the screen.
Which of my students are stuck on osmosis, by section?
37 of your 142 students mixed up the direction of osmosis this week (58 questions, October 14–20): L01 11 and L02 9 with Priya Natarajan, L03 10 and L04 7 with Tomás Ferreira. Most think the salt moves into the cell.
Draft the 10-minute re-teach from my slides
Drafted only from your materials: Figure 4.2, then the red blood cell example from Lecture 7 slides 12–15, two ungraded check questions (checked against your sealed keys, highest match 0.22), and a one-line exit question, ready for your approval.
Why was the worked example rewritten last night?
Lucía M.'s question matched graded Lab 4 post-lab question 2 at 0.91. The Tutor's first draft of the worked example used the onion cell, the graded case, so the Guardrail agent sent it back. The rewrite uses a red blood cell in 3% saline, and Lucía wrote her own answer.
Which questions had no source this week?
11 questions, all on the sodium–potassium pump. Lab 5 handout page 4 is a scanned image with no text, so the Tutor told the students it could not find it rather than guess. I can read the page with text recognition for you to confirm, then draft answers to all 11.
13 screens from the working solution, on its sample data. Pick one to see it large.
Every rule and threshold the agents work to is a setting: a change makes a new version, and it is on the audit trail.
| Setting | Default | Choose from |
|---|---|---|
| Reading level (BIO 141) | Grade 10 | College ceiling grade 12; instructors may set it lower, never higher |
| Worked example | Only after the student tries once | Or off: the ladder ends at Hint 2 |
| Exam window | Tutor off during Midterm 2, Wed, Oct 28, 8:00–9:15 a.m. | Students see when it is back |
| Source match to answer (Tutor agent) | 0.78 | Below it, the student is told it is not in the course materials |
| Graded-key match to rewrite (Guardrail agent) | 0.70 | Overrides only after the graded item closes, with a reason code |
| Translation confidence (Language agent) | 0.90 | Below it, a bilingual person reads the reply within 24 hours |
| Safety alert acknowledged | 15 minutes | 5–30 minutes; then the Director of Counseling, and campus safety for threats |
| Conversations kept | Until one year after the term ends | Deleted within 30 days when a student opts out, unless a legal hold applies |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
Sources: Student Voice survey, 2025 (college students' use of AI, many as a “24/7 tutor”) · a publisher's AI tutor usage report, December 2024 (heaviest use 6 p.m.–3 a.m. on weeknights) · FERPA, 34 CFR Part 99 · ADA Title II web rule, 28 CFR Part 35, Subpart H · “target” = design goal, measured in the live solution · “demo” = seen in the working solution, on its sample data · “typical” = published figures · “estimated” = our arithmetic, kept modest. Brandvale Community College and the people named on this page are fictional: characters and sample data in the working solution.
An AI teaching assistant helps students with their coursework when the instructor is not available, ideally from the instructor's own materials and within the instructor's rules. AI Teaching Assistant gives students cited answers from their own course materials: a Tutor agent replies as two hints and a worked example, each cited to a page, a separate Guardrail agent checks every rung against sealed answer keys, and each morning the instructor sees what her sections are stuck on and approves what to re-teach.
A general chatbot answers from the open web, cites nothing a student can check, and will often hand over the answer to graded work. This AI tutor answers only from what the instructor uploaded (the syllabus, lab handouts, slides and readings), cites the page for every hint, and when no page answers the question it says so and points to office hours instead of guessing.
No. Answer keys are uploaded sealed: the Tutor agent never reads them, and the Guardrail agent checks every draft against them before a student sees it. When a question matches a graded item, the student gets hints and a worked example on a different case, which opens only after the student tries once, so the student still writes the answer.
Yes. The Language agent replies in the student's language, in 12 reply languages, and keeps the English beside each message for the instructor with its translation confidence. Below 0.90, a bilingual learning assistant reads the reply within 24 hours.
The Safety agent shows the student the 988 Suicide & Crisis Lifeline and the after-hours counseling line at once, and alerts the on-call counselor with the flagged exchange only, on a 15-minute acknowledgment clock that escalates to the Director of Counseling. It never diagnoses or contacts the student, and the instructor sees only that a referral happened.
Conversations are treated as education records. The privacy settings lock on no training on student data and zero retention at the model provider, keep conversations until one year after the term ends, delete them within 30 days when a student opts out, and hide names from anyone outside the course. Students see a plain notice in English and Spanish, and every view of a conversation is logged.
Yes. The agents draft, check and alert; they never send to a student on a person's behalf, approve, or change a record. The instructor approves every re-teach and every note a learning assistant drafts, can tighten her course's rules but never loosen the college floor, and can override a rewrite only with a reason code. College-wide changes need the Dean and the Vice President of Instruction.
The Agentic Solution Engine builds and deploys it from your requirements and documents (your college's guardrails, a gateway course's materials and your counseling protocol), and it goes live once every quality gate has passed. We will walk you through it on one of your own courses first.
Every course file fixed before April 2027, and heard right by a screen-reader test.
1-min demoAcademic Integrity Case ManagementReferrals screened on how the work was written, with both sides in one packet.
1-min demoLearning Outcomes AssessmentAssessment reports built from scored student work, where every number opens its papers.
1-min demoWe'll run AI Teaching Assistant on a gateway course's syllabus, handouts and slides, with your rules.